[LLVMdev] git

[David A. Greene]

Chris Lattner <clattner at apple.com> writes:

> On Jul 28, 2011, at 2:16 PM, David A. Greene wrote:

> Disagreed.  The point is that I should not see a stream of 20
> decomposed patches from you.  When I get to one that is "wrong" or
> needs changes (e.g. patch 6), then all the other patches after it get
> ignored.  This is silly.

It is silly.  I see no reason to simply ignore the later patches unless
patch 6 needs a substantial rework.

> Instead, what I should see are (for example) and be asked to review
> them.

Oops, I think something got botched here.  I can't parse the above.  :)

> Here's a hypothetical example of what I want:
>
> Send in or directly commit 5 patches if they are "obvious" cleanups.
> These are patches that just move things around in obvious ways, fix
> formatting, whatever.

Ok.

> Send in a patch 6 and wait for review of *just it* saying "this does
> something crazy, here is why, here is why it is the best thing to do".
> Someone will review it, and if there are any problems, it goes back to
> you for revision.  It eventually goes in when it is good for the tree.

Ok.

> After that patch is in, another stream of obvious patches are unblocked and go directly in.

Yep.

> Patch 12 comes around, gets reviewed just like #6... etc.
>
> I *don't* want to see all 20 patches up front.

Ok, now I'm understanding you better.

> No.  The fundamental disagreement/misunderstanding here is that you
> are optimizing for an out-of-tree maintainer pushing "batches" of work
> upstream.  This is not what I'm at all interesting in optimizing for,
> and if you're irritated about the turn-around time that it takes to
> get review, then you shouldn't be optimizing for this either.

Thanks, that clarifies things.  So you don't so much demand individual
cherry-picks to master as you demand that merges to master be smallish
and incremental.  I think that is entirely reasonable.

With my most recent patch, I submitted something I considered logically
connected.  I was asked to rework it and break it up.  I did that and
submitted the resulting patch stream for review.  No one seemed to
complain about that so I wasn't grokking your dislike of that apporoach.
Now I am much more clear about how to use this git thing with LLVM.  :)

                                    -Dave